Most DBAs never think about the internal structure of a Data Pump dump file until something breaks.
The export finishes successfully. The dump file exists. Object storage upload completes. Then suddenly an import job fails halfway through a cloud migration because the dump file metadata cannot be interpreted properly by the target workflow. Or worse, a multi-terabyte export streaming directly into cloud storage stalls because the metadata handling does not behave well with object-based writes.
That is exactly the kind of operational problem Oracle quietly addressed in Oracle AI Database 26ai.